PEIRA : Learning Predictive Encoders through Inter-View Regressor Alignment
Michael ArbelUniv. Grenoble Alpes, InriaCNRS, Grenoble INP, LJK [email protected], Basile TerverEcole Normale Supérieure / PSLInria Paris [email protected] Jean PonceEcole Normale Supérieure / PSLNew York University [email protected] · 2026
Read PEIRA : Learning Predictive Encoders through Inter-View Regressor Alignment by Michael ArbelUniv. Grenoble Alpes, InriaCNRS, Grenoble INP, LJK [email protected], Basile TerverEcole Normale Supérieure / PSLInria Paris [email protected] Jean PonceEcole Normale Supérieure / PSLNew York University [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-17671.txt.
Description
Michael Arbel, Basile Terver, and Jean Ponce analyze a JEPA variant, a self-supervised method predicting one data view from another. Popular methods like SimSiam and BYOL train teacher-student pairs without a clear objective. The trio characterize stability for a regularized linear regressor: sound equilibria align with leading canonical correlation subspaces, while collapsed ones can also attract. Their PEIRA method, built on the optimal regressor trace, has only sound global minimizers as stable equilibria. Tests on ImageNet-1K and CIFAR-10 match VICReg and LeJEPA baselines.
Catalog record
| CREATOR | Michael ArbelUniv. Grenoble Alpes, InriaCNRS, Grenoble INP, LJK [email protected], Basile TerverEcole Normale Supérieure / PSLInria Paris [email protected] Jean PonceEcole Normale Supérieure / PSLNew York University [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 190950 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.17671